---
license: cc-by-4.0
tags:
- pytorch
- computer-vision
- remote-sensing
- mars
- dem-prediction
- u-net
- multi-task-learning
datasets:
- ESA-Datalabs/MCTED
---

# MarsDEMNet

MarsDEMNet is a comparative deep learning study for single-image Digital Elevation Model (DEM) prediction from Mars CTX satellite imagery. Four architectures are evaluated, a classical Random Forest baseline, a single-output U-Net, a multi-output U-Net with multi-task learning, and an encoder depth ablation — all trained on the MCTED dataset of 80,898 paired CTX orthoimage and DEM patches.

## Model Details

### Model Description

MarsDEMNet addresses a fundamental coverage asymmetry on Mars: while the CTX instrument has photographed ~99.5% of the Martian surface at 5–6 m/pixel, high-resolution stereo DEMs exist for only ~0.5–1% of that coverage. Models trained on MCTED learn to predict dense elevation maps from single optical images, extending effective DEM coverage to nearly the entire planet.

- **Model type:** Convolutional encoder-decoder (U-Net)
- **License:** CC-BY 4.0
- **Finetuned from:** Trained from scratch — no pretrained weights

### Model Sources

- **Repository:** https://github.com/harshithkethavath/MarsDEMNet
- **Dataset:** https://huggingface.co/datasets/ESA-Datalabs/MCTED

## Checkpoints

Four model checkpoints are provided:

| File | Architecture | Val RMSE | Val MAE | Delta-1 |
|---|---|---|---|---|
| `marsdеmnet-unet-elevation-4block.pt` | Single-output U-Net, 4-block encoder, 7.8M params | 74.38m | 52.86m | 0.418 |
| `marsdеmnet-unet-multitask-4block.pt` | Multi-output U-Net, 4-block encoder, 7.8M params | 74.29m | 52.68m | 0.422 |
| `marsdеmnet-unet-multitask-3block.pt` | Multi-output U-Net, 3-block encoder, 1.9M params | 82.80m | 58.29m | 0.440 |
| `marsdеmnet-unet-multitask-5block.pt` | Multi-output U-Net, 5-block encoder, 31.4M params | 59.88m | 42.67m | 0.409 |

The 5-block multi-output model is the best overall, achieving 19% lower RMSE than the 4-block baseline with no overfitting observed.

## How to Get Started

```python
import torch
from scripts.deeplearning.unet import UNet

# Single-output (elevation only) — 4-block
model = UNet(in_channels=1, out_channels=1, num_blocks=4, base_ch=32)
ckpt  = torch.load("marsdеmnet-unet-elevation-4block.pt", map_location="cpu")
model.load_state_dict(ckpt["model_state"])
model.eval()

# Multi-output (elevation + slope + roughness) — 5-block (best)
model = UNet(in_channels=1, out_channels=3, num_blocks=5, base_ch=32)
ckpt  = torch.load("marsdеmnet-unet-multitask-5block.pt", map_location="cpu")
model.load_state_dict(ckpt["model_state"])
model.eval()

# Inference
with torch.no_grad():
    # optical: (1, 1, 518, 518) normalized CTX patch
    pred = model(optical)
    # Single-output: pred shape (1, 1, 518, 518) — elevation
    # Multi-output:  pred shape (1, 3, 518, 518) — [elevation, slope, roughness]
```

Input normalization: clip to 2nd–98th percentile, then z-score per patch. DEM targets are mean-subtracted per patch (relative elevation in meters).

## Training Details

### Training Data

MCTED (Mars CTX Terrain-Elevation Dataset) — 80,898 paired CTX orthoimage and DEM patches derived from 1,122 quality-filtered stereo scenes. Geography-aware train/val split at the scene level to prevent spatial leakage. Train: 65,090 patches. Val: 15,808 patches.

### Training Procedure

- **Optimizer:** AdamW, lr=1e-4, weight_decay=1e-4
- **Schedule:** Cosine annealing to 1e-6 over 50 epochs
- **Early stopping:** Patience 10 on val RMSE
- **Batch size:** 16
- **Augmentation:** Random horizontal/vertical flips and 90° rotations applied jointly to image and labels
- **Loss:** Masked MAE (single-output); weighted sum of masked MAE losses (multi-output, uniform 1:1:1 weights)
- **Training regime:** fp32
- **Hardware:** NVIDIA H100 GPU

### Preprocessing

- CTX patches: percentile clip (2nd–98th) + per-patch z-score normalization
- DEM patches: per-patch mean subtraction (relative elevation)
- Validity masking: logical AND of NaN mask and deviation mask; invalid pixels excluded from loss and metrics

## Evaluation

### Metrics

- **MAE** — mean absolute elevation error in meters
- **RMSE** — primary ranking metric; penalizes large errors
- **Delta-1** — fraction of valid pixels where max(pred/gt, gt/pred) < 1.25

### Results

| Model | Params | Val RMSE | Val MAE | Delta-1 |
|---|---|---|---|---|
| Random Forest (classical baseline) | — | 58.39m (elev std) | 41.29m | — |
| Single-output U-Net (4-block) | 7.8M | 74.38m | 52.86m | 0.418 |
| Multi-output U-Net uniform (4-block) | 7.8M | 74.29m | 52.68m | 0.422 |
| Multi-output U-Net (3-block ablation) | 1.9M | 82.80m | 58.29m | 0.440 |
| Multi-output U-Net (5-block ablation) | 31.4M | **59.88m** | **42.67m** | 0.409 |

## Bias, Risks, and Limitations

- Models are trained on regions of Mars where stereo DEMs exist, which are geographically biased toward scientifically interesting terrain. Performance on flat, featureless plains may be lower.
- Textureless terrain with no illumination gradient provides no depth cue, a known failure mode.
- Predictions are relative elevation (mean-subtracted per patch), not absolute MOLA-referenced altitude.
- Not suitable for safety-critical mission planning without further validation.

## Technical Specifications

### Model Architecture

U-Net encoder-decoder with configurable depth. Each encoder block: Conv2d(3×3) → BatchNorm → ReLU × 2 → MaxPool. Decoder: bilinear upsampling + lateral skip connections. Multi-output variant has three separate 1×1 conv heads for elevation, slope, and roughness.

## Citation

If you use MarsDEMNet, please cite:

```bibtex
@misc{marsdеmnet2026,
  title     = {MarsDEMNet: Classical and Deep Learning Approaches for Single-Image Digital Elevation Model Prediction from Mars CTX Imagery},
  author    = {Harshith Kethavath},
  year      = {2026},
  publisher = {GitHub},
  url       = {https://github.com/harshithkethavath/MarsDEMNet}
}
```